Skip to main content
RankRift
Menu

Competitive intelligence from AI answers

AI answers are market research. This hub covers who owns the default recommendation in a category, which sources decide it, how category framing changes whether you appear at all, and how to benchmark competitors fairly.

Browse the prompt-set patterns · Use the manual snapshot method

In short

This hub reframes AI-visibility measurement as market research rather than a scoreboard: which brand an assistant names first when a buyer asks for the best option in a category, which sources that answer leans on, whether your category is even framed the way your buyers describe it, and how the language of an answer reveals how the market publicly describes you. It also sets the rules for benchmarking against competitors fairly - identical prompts, identical run counts, identical time windows, published deltas - and the line between a factual statement about model output and a quality judgement.

Who this is for

  • SEO and organic growth leads who need to show leadership who currently owns the default recommendation in their category
  • Agency strategists running competitive teardowns as part of a pitch or a quarterly review
  • Founders and heads of marketing who keep hearing a competitor's name come back from prospects who asked an assistant
  • Product marketers testing whether the category name they use matches the one buyers actually say
  • PR and comms leads deciding which publications, communities, and review sites are worth pursuing

When to use this

  • You know a competitor gets recommended and you want to understand what the answer is built from
  • You are choosing between category framings and want evidence rather than a workshop opinion
  • Someone has shown you a competitor comparison chart and you need to check whether it was fair
  • You are planning earned coverage and want to target the sources that actually appear in your category's answers
  • A competitor's presence in answers changed and you need to work out whether they moved or the model did

The question is not "where do we rank" - it is "who is the default"

A results page shows ten options and lets a buyer choose. An assistant asked for the best option in a category often returns two or three names and a reason. That is closer to a shortlist handed over by a colleague than to a ranking, and it means the competitive question changes shape: not what position you hold, but whether you are in the set at all, and what the answer says about you when you are.

Treated that way, a prompt set becomes a research instrument. Run it repeatedly across assistants and you get a picture of which brands the market's public record has made legible enough to recommend, which is a different and more useful thing than a traffic estimate.

It is also a picture with error bars. Answers vary by model, phrasing, region, personalization, and date, so one screenshot of one answer is a story, not a finding.

Which sources decide your category

Assistants that retrieve before answering are drawing on a small number of pages per category, and those pages are rarely the ones a brand assumes. In some categories it is trade press; in others it is a review platform, a documentation site, a forum thread, or a single well-structured comparison article that happens to be easy to quote.

The useful exercise is not learning which sources are generally influential. It is identifying which sources appear behind your prompts, in your category, on the dates you ran them - and then deciding which of those are legitimately reachable through earned coverage, better documentation, or accurate third-party listings.

That mapping is also the honest version of a link strategy conversation: you are trying to be accurately described in the places a retriever already reads, not trying to manufacture presence.

Category framing decides whether you appear at all

If buyers ask for the best option in a category you do not describe yourself as belonging to, no amount of measurement will rescue the result. You are not losing the comparison; you are absent from it.

This is testable. Write the same buyer intent three or four different ways - the term you use internally, the term your competitors use, the term an outsider would use, and the job-to-be-done phrasing - then run each as its own prompt group and compare who appears in each. A brand that shows up strongly under one framing and vanishes under another has a positioning problem, not a visibility problem.

Read the wording of the answers too. An assistant's description of a category is a compressed synthesis of how the market publicly describes it. If the summary of your product sounds like a competitor's marketing, that is what the public record currently says about you.

Test alternative framings as separate prompt groups with their own run history. Mixing framings into one score produces a number that averages away the finding.

Fair benchmarking rules

Most competitor charts in this category are unfair by accident. Someone runs their own brand's prompt set thoroughly and a competitor's casually, or compares this month's runs against a rival's numbers from a different model version, and the resulting bar chart flatters whoever built it.

Apply these rules to any comparison you build yourself and to any chart an agency or vendor shows you.

A measured visibility figure is a factual statement about what a model output contained on a given date. It is never a judgement about whether a product is good.

Publishing competitor results without doing harm

If you publish a comparison that names other brands - in a report, a pitch deck, or a public post - you have taken on an obligation. Someone measured unfavourably deserves to know how the number was produced and to have a route to challenge it.

Use a simple rule: publish the method, carry the raw prompt set and run dates with the figure, provide a re-run record, and date corrections instead of quietly editing them. Apply the same standard to competitor numbers.

The framing matters as much as the process. "Brand X was named in a given share of runs for this prompt set on these dates" is defensible. "Brand X is losing" is a claim about a business you have not measured.

The rest of this cluster

Start with the teardown if you need a result this week, or with the prompt-set patterns if you would rather not write sixty prompts from scratch.

Frequently asked questions

How is this different from measuring my own AI visibility?

Same instrument, different question. Measuring your own visibility tells you whether you appear and how often. Competitive intelligence asks who owns the default recommendation in the category, which sources that answer draws on, and what the wording reveals about how the market describes each player. The measurement pillar owns how you count; this pillar owns what you count it against and what the pattern means for positioning. Most teams need both, but they answer to different meetings.

How many competitors should I track?

Fewer than you want to. Every additional brand multiplies runs, and run volume is the real cost of this work. A practical starting point is the two or three names that actually come back from prospects, plus any brand that keeps appearing in your prompt results that you did not expect. Add brands when a specific decision depends on them. A tracked set you can run at full sample size beats a longer list you can only sample once.

A competitor appeared out of nowhere in answers. What happened?

Four candidates, and you cannot skip the elimination: a provider shipped a model or retrieval change, a source the assistant leans on published something new about them, they genuinely changed their public footprint, or your own runs moved within normal variance and you are reading noise. Check the boring one first by looking at your run history spread. Then check whether the shift shows up across multiple assistants, which points at their side rather than one provider's.

No, and this is worth saying out loud in any deck that carries these numbers. A visibility measurement records what a model output contained on given dates for a given prompt set. It reflects how legible and how well-documented a brand is across the sources a retriever reads, not product quality, customer outcomes, or fit for your buyer. Treating it as a quality verdict is how these charts become unfair, and it is the fastest way to lose credibility with a skeptical executive.

Can I use these results in a sales deck or a public post?

You can, if you carry the method with the number. Publish the prompt set, model versions, run dates, sample size, and spread alongside the chart, phrase findings as statements about model output rather than about the competitor's business, and give named brands a route to ask for a re-run. Without those, you are publishing an unfalsifiable claim about someone else's company, which is a reputational and possibly legal exposure that a bar chart is not worth.

What if my competitor's brand name is a common word?

Decide the counting rule before you run, write it down, and apply it identically to every brand including your own. Common options are requiring the brand name to appear in a recommendation context, requiring a linked or cited source that resolves to the company, or excluding ambiguous matches entirely and reporting the exclusion rate. Any of these is defensible; changing the rule after seeing results is not. Ambiguity handling belongs in your method notes, not in a footnote nobody reads.

Where do I get prompts without writing sixty of them?

Start from the buyer-stage prompt patterns and edit them. They give you buyer-stage coverage - discovery, comparison, shortlist, objection - so you are less likely to write only prompts your own brand happens to answer well. Then add the phrasings your sales team actually hears, which are usually blunter than anything a marketer writes. Version and date whatever you end up with, because a prompt set that quietly changes makes every trend line built on it meaningless.

Start from a real prompt set

Build a category prompt set from the patterns, run it across your brand and two competitors under identical conditions, and you have a defensible teardown by the end of the week.

Open the prompt-set patterns · See the research method · Use the briefing template